CoHuB: A Collaborative Humanoid Loco-Manipulation Benchmark

Humanoid robots are well suited to diverse physical tasks in human environments, yet many such tasks call for collaboration. Whether assisting a partner or jointly manipulating a shared object, two humanoids must coordinate their actions toward a shared goal. However, existing benchmarks do not provide a common setting for evaluating how humanoid robots physically collaborate under egocentric observations. We introduce CoHuB, a simulation benchmark for collaborative loco-manipulation with full-body humanoid robots and dexterous hands. CoHuB provides a diverse suite of collaborative tasks that vary locomotion and physical coupling in a controlled way, together with synchronized demonstrations collected through a dual-operator VR teleoperation pipeline, in which each operator controls one robot from its egocentric view. Our results across centralized and decentralized settings highlight the capabilities required for successful collaborative loco-manipulation, providing guidance for future collaborative humanoid policy learning. Code and data will be released.